Mistral has released the open weights for Mistral Large 4, its 1 trillion-parameter model, on Hugging Face Inc. on October 29, claiming the model now matches or edges the best open-weight competitors on agentic coding and cyber-defense benchmarks after three additional weeks of reinforcement learning since its October 6 preview1,4,5.
The model was trained on 3,800 Nvidia Corp. NVIDIA Grace Blackwell GPUs in Mistral's European data centers3. It activates 52 billion of its 1 trillion parameters per task, handles text, image, and code inputs, and offers a one-million-token context window. Mistral is also shipping a quantized version that runs on four NVIDIA B200 or NVIDIA B300 GPUs.
Three weeks of reinforcement learning
Pierre Stock, Mistral's Vice President of Science, said the open-weight edition continues reinforcement learning from the preview and delivers substantial gains with no signs of saturation. On Terminal-Bench 4.0, which tests agentic coding inside a terminal environment, the model's accuracy climbed from 28% at preview to 41%, placing it just ahead of GLM-5.3 at 40%. Mistral told Fortune the two models tie on that benchmark6.
On Artificial Analysis' Coding Agent Index, an independent benchmark measuring full software-engineering task completion, Mistral said Large 4 narrowly tops all other open models7. The company added that Large 4 uses 47% fewer tokens per task on average than GLM-5.3, a gap that compounds during long-running agent sessions.
Stock told reporters that Mistral fully supports open weights, calling them essential for business continuity, transparency, and auditability2. Open weights allow companies to deploy the model on their own hardware, within their own firewalls, without exposing sensitive data to external parties.
Cyber-defense scores
In cyber defense, Mistral Large 4 scored 79% on CyberGym-E2E, a test that asks models to find, reproduce, and patch real memory-safety bugs in open-source C and C++ projects. That result matches MiMo V2.6 Pro and beats DeepSeek-V4.1-Flash at 76%.
The model ships under a modified MIT license that permits commercial use but requires a separate commercial license for anyone reselling the model as a service.
Reasoning controls and sovereign deployment
Early testers, described by Mistral only as governments and enterprises, requested more control over the model's reasoning depth. In response, Mistral added low and medium reasoning modes alongside the existing high mode and made the high mode 50% more efficient. Stock said the team incorporated customer feedback within the same reinforcement learning run, adding the two reasoning modes while improving efficiency and customizability.
Mistral framed the release around sovereign AI deployment, positioning open weights as a way for European enterprises and governments to possess and control frontier cyber-defense capabilities without relying on U.S. or Chinese cloud providers.
ANALYSIS The open-weight release puts Mistral Large 4 in direct competition with GLM-5.3 across the same benchmark suite, with the token-efficiency claim offering a cost argument even where raw accuracy is close to parity. The sovereign-deployment pitch adds a procurement angle distinct from pure performance.